Академический Документы
Профессиональный Документы
Культура Документы
and
J. Nelson Wright
Acuson, 1220 Charleston Rd., Mountain View, CA 94039-7393, USA
Abstract
A detailed mathematical investigation of multi-bit quantizing systems using
non-subtractive dither is presented. It is shown that by the use of dither having
a suitably-chosen probability density function, moments of the total error can be
made independent of the system input signal, but that statistical independence
of the error and the input signals is not achievable. Similarly, it is demonstrated
that values of the total error signal cannot generally be rendered statistically
independent of one another, but that their joint moments can be controlled and
that, in particular, the error sequence can be rendered spectrally white. The
properties of some practical dither signals are explored and recommendations are
made for dithering in audio, video and measurement applications. Some of the
results presented here are known to a handful of individuals in the engineering
community, but many appear to be unpublished. In view of many widespread
misunderstandings regarding non-subtractive dither, formal presentation of these
results is long overdue.
Introduction
Q(w)
Q(w)
(a)
(b)
1.1
Q(w) =
+
2
for a mid-riser quantizer, where the floor operator, b c, returns the greatest integer
less than or equal to its argument. The step size, , is commonly referred to as a
LSB (least significant bit), since a change in input signal level of one step width
corresponds to a change in the LSB of binary coded output. Throughout the sequel,
quantizers of the mid-tread variety will be assumed, but all derived results have obvious
analogs for mid-riser quantizers and all stated theorems are valid for both types.
Quantization or requantization introduces an error signal, q, into the digital data
stream, which is simply the difference between the output of the quantizer and its
input:
4
q(w) = Q(w) w,
4
(1)
q(w)
[LSB]
0.5
-1.0
1.0
w
[LSB]
-0.5
(q) =
1 , < q ,
2
2
otherwise.
0,
m 4
E[q ] =
q m pq (q)dq,
E[f ] =
f (q)pq (q)dq.
The zeroth moment of any random process (i.e., E[q 0 ]) is identically equal to unity.
The first moment is usually referred to as the mean of the process, whereas the term
variance refers to the quantity E (q E[q])2 = E[q 2 ] E 2 [q]. It is clear that if the
mean of a random process is zero, its variance and second moment are equal.
6
If a quantization error signal is distributed according to Eq. (1), its moments are:
E[q] = 0
2
E[q 2 ] =
12
E[q ] =
(2)
(3)
1
m+1
0,
m
, for m even,
(4)
for m odd.
Eq. (3) is the familiar expression for the variance of the quantization error in the
classical model.
The CMQ is valid for input signals which exhibit smooth pdfs and which are large
relative to an LSB [31, 25]. It fails catastrophically for small signals and many particularly simple (e.g., sinusoidal) signals, for which the quantization error retains the
character of input-dependent distortion, rather than noise. The mid-tread quantization of a small signal of peak amplitude less than 0.5 LSB provides a simple example
of this failure: the quantizer output is null, and the quantization error is just the input
sign-inverted. Such an error is not uniformly distributed, iid or independent of the
input. In such cases, application of an appropriate dither can be used to temper the
statistical properties of the error signal.
1.2
Schematics of subtractively dithered and non-subtractively dithered quantizing systems are shown in Fig. 3. In each case we denote the system input by x and the
system output by y. We thus distinguish the system input from the quantizer input,
which we continue to denote by w and which is given by w = x + . represents
the dither signal, a strict-sense stationary random process which is assumed to be
statistically independent of x. Similarly, the total error of each quantizing system is
defined as the difference between the system output and system input, and is denoted
7
dither ,
+
input
x
w=x+
quantizer
Q
(a) subtractively
dithered
output
y = Q(w) -
= x + q(x+)
=x+
dither ,
+
input
x
w=x+
output
y = Q(w)
= x + + q(x+)
=x+
quantizer
Q
(b) non-subtractively
dithered
channel
Figure 3: Dither quantizing systems: (a) subtractively dithered (SD), (b) nonsubtractively dithered (NSD).
by
4
=yx
to distinguish it from the quantizer error, q = Q(w) w.
The total errors introduced by subtractively dithered and non-subtractively dithered
systems are not identical. In a subtractively dithered system, the dither is subtracted
from the quantizer output to yield the system output. Hence, for such a system:
= yx
= Q(x + ) (x + )
= q(x + ).
On the other hand, for the non-subtractively dithered system:
= yx
= Q(x + ) x
= q(x + ) + .
In neither case is the total error equal to q(x) as in an undithered system (i.e., one for
which 0), although in an SD system the total error does equal the quantization
error associated with the total quantizer input w.
It has been shown by Schuchman [3] that the total error induced by an SD quantizing system can be rendered uniformly distributed for arbitrary input distributions
if and only if the dithers characteristic function or cf (the Fourier transform of its
pdf [32, 33]) obeys a certain condition. Defining the Fourier transform operator, F[ ],
by
4
f (x)ej2ux dx,
and denoting the dither pdf and cf as p () and P (u), respectively, Schuchmans
condition is that
P
=0
k Z0 ,
(5)
where we take this opportunity to define the set Zn0 as the set of all n-vectors with
integer components with the exception of the zero vector 0 = (0, 0, . . . , 0); i.e., Zn0 =
Zn \ 0.
Furthermore, it can be shown [8, 4, 26] that the total error in an SD quantizing
system is statistically independent of the system input if and only if Eq. (5) holds.
Thus, dither obeying Schuchmans condition renders the error statistically independent
of the input and uniformly distributed. In particular, it exhibits a variance of 2 /12.
In these regards, then, it resembles the idealized quantization error of the CMQ. The
simplest random process satisfying Schuchmans condition is one exhibiting a uniform
pdf
p () = (),
whose associated characteristic function is a sinc function:
4
sin(u)
.
u
(Different authors employ slightly different definitions of the sinc function. We will
retain the above throughout the sequel.)
It can also be shown [8, 4, 26] that subtractive dither will render distinct samples
of the total error signal statistically independent of one another for arbitrary input
distributions if and only if
P1 ,2
k1 k2
,
=0
(k1 , k2 ) Z20 ,
(6)
transform of their joint pdf, p1 ,2 (1 , 2 )). This condition is satisfied by any dither
which is iid, so that P1 ,2 (u1 , u2 ) = P (u1 )P (u2 ), and which satisfies Eq. (5). For
instance, this means that a subtractively dithered quantizing system employing iid
dither of uniform distribution produces an iid total error signal whose values are uniformly distributed and statistically independent of the input. The error thus behaves
like a purely additive independent noise process, as postulated by the CMQ. This
beautiful result represents the ideal outcome for a quantization operation.
Unfortunately, subtractive dithering is difficult to use in many practical systems
since the dither signal must be available at each end of the channel. This requires
either the transmission of the dither values or the use of synchronized noise sources
(pseudo-random number generators) separated, in general, by both time and distance.
Furthermore, any digital processing of the dithered signal would necessitate processing
of the dither prior to subtraction. For reasons such as these, the possibility of using
dither without subsequently subtracting it is frequently of interest.
We will see that non-subtractively dithered systems, as distinct from subtractively
dithered ones, cannot render the total error statistically independent of the input.
Neither can they make temporally separated values of the total error statistically
independent of one another. However, we shall prove that they can render any desired
statistical moments of the error signal independent of the input and regulate the joint
moments of errors which are separated in time. The theory underlying these features of
non-subtractive dither is developed in Section 2, and is subsequently used to explore
the properties of some practical dither signals in Section 3. Section 4 explores the
important special case of quantizing systems in which the available dither is discrete
valued, while Section 5 summarizes the most important observations and conclusions.
11
We begin by describing the relationship between the total error and the input signal
in probabilistic terms.
2.1
The dependence of the total error on the system input can be analyzed in terms of
its pdf as a function of a specified input value. This function is referred to as the
conditional pdf , or cpdf , of the total error and is denoted p|x (, x) throughout the
following discussion.
In order to derive an expression for p|x (, x), we consider a non-subtractively
dithered quantizing system, as in Fig. 3(b), with a specified system input value, x.
The input to the quantizer is w = x + , the sum of the system input and the
statistically independent dither process. This sum has a cpdf
pw|x (w, x) = p (w x).
Fig. 4 shows that total error depends not only on the system input value, but also
on the value of the dither. In particular, if the input to the quantizer, w, is between
/2 and +/2, the output will be nil (for a mid-tread characteristic) so that the
error is = x. Similarly, if the input to the quantizer is between +/2 and +3/2
the output will be +, so that = x + . Hence, the pdf of the error for a fixed
input is a series of delta functions separated by intervals of , each weighted by the
probability that w falls upon the corresponding quantizer step:
p|x (, x) =
( + x k)
k=
12
+k
2
+k
2
p (w x)dw.
y
p (w, x)= p ( w-x)
w|x
w = x +
input, x
Figure 4: Cpdf of the quantizer input showing its justification relative to the quantizer
transfer characteristic.
13
In the parlance of Widrow [7], the error cpdf is an area sampled version of the quantizer
input cpdf.
Writing the integral in the last equation as a convolution (denoted by ?) of p with
a rectangular window function, , it reduces to
p|x (, x) = [ ? p ]()W ( + x),
where
4
W () =
(7)
( k)
k=
is a train of Dirac delta functions separated by intervals of width .1 Thus the pdf of
is given by
p () =
= [ ? p ]()[W ? px ]().
(8)
A problem arises in the formalism if the dither is null (p () = ()) and the system input occurs
at a quantizer step edge, since the product of the generalized functions W ( (2n + 1)/2) and
() is not conventionally defined. It is shown in [25] that an appropriate definition of this product
for the purposes at hand is
1
[(
2
n) + ( (n + 1))].
14
Proof : Eq. (7) makes it clear that p|x (, x) cannot be rendered independent of x by
any choice of dither pdf, since the convolution of any dither pdf (which must be nonnegative everywhere) with a rectangular window function yields a function at least as
wide as the rectangular window. Hence, at least one delta function always makes a
contribution to the sum, and the position of that delta function is dependent on the
system input.
Taking the Fourier transform of Eq. (8) we find that the characteristic function of
is given by
h
k
k
k
=
sinc u
P u
Px
,
k=
(9)
where Px is the arbitrary cf of the input signal and P is the cf of the dither. In order
for to be uniformly distributed, this must reduce to sinc (u) for some choice of P .
Suppose that this is possible, in which case we obtain
sinc (u) =
sinc u
k=
k
k
k
P u
Px
.
= 0 = Px
which contradicts the assumption that Px is arbitrary. Thus the total error cannot
be made uniformly distributed in a non-subtractively dithered system for inputs of
arbitrary distribution.
2
The counterintuitive nature of this result is the source of much confusion regarding
NSD systems. For instance, it is tempting to accept the following line of reasoning:
15
suppose that a dither satisfying Schuchmans condition (Eq. (5)) is used so that q is
independent of x. Then, since is also independent of x, the total error = q + is the
sum of two random processes both of which are independent of x and thus should be
independent of x as well. This conclusion is flatly false. In an NSD quantizing system,
given the value of q + we know that the possible values of x satisfy the equation
x = (q + ) + k, k Z, so that the distribution of x is highly dependent on q + .
To elucidate the source of the problem we may reason as follows: for arbitrary random
variables q, , and x and a fourth = q + (none of these necessarily representing
quantities in a quantizing system) it is clear that
p|q,,x(, q, , x) = ( q ).
Then
p,x (, x) =
=
pq,,x ( , , x)d,
Z Z Z
= Pq,,x (u , u , ux ).
By definition, and x are statistically independent of one another if and only if
P,x (u , ux ) can be written as a product of two functions, one involving u alone while
the other involves ux alone. From the above we see that this is the case if and only if
Pq,,x (u , u , ux ) = Pq, (u , u )Px (ux ).
16
Unfortunately, knowing as we do that Pq,x (uq , ux ) = Pq (uq )Px (ux ) and P,x (u , ux ) =
P (u )Px (ux ) is simply not sufficient to ensure satisfaction of the latter condition. Of
course, the result would hold if {q, , x} formed a set of independent random variables;
that is, if it were the case that
Pq,,x (uq , u , ux ) = Pq (uq )P (u )Px (ux ).
However, this even stronger condition is certainly not met in an NSD quantizing system
with an arbitrarily distributed input.
Since we have shown that statistical independence of the total error from the system
input is not achievable, we now turn our attention to the possibility of controlling
moments of the error. For many applications, controlling relevant error moments is
just as good as having full statistical independence of the input and error processes.
2.2
m p ()d.
It can be shown that these moments may also be expressed in terms of the cf of the
given random variable as [33]:
E[m ] =
(m)
where P
m
j
2
P(m) (0),
(10)
j
2
m X
G(m)
k=
17
k
Px
k
,
(11)
where
4
(12)
Since the cf, Px , of the system input is arbitrary we obtain the following result [26]:
Theorem 2 In an NSD quantizing system, E[m ] is independent of the distribution
of the system input, x, if and only if
G(m)
=0
k Z0 .
(13)
E[ ] =
j
2
m
G(m)
(0),
which is precisely the m-th moment of a notional random process with cf G and pdf
? p , although this is not, of course, the pdf of . We can derive the following
expressions for the moments of the total error in terms of the moments of the dither
signal by direct differentiation of G (u):
E[] = E[]
(14)
E[2 ] = E[ 2 ] +
m
E[ ] =
m
bX
2 c
`=0
m
2`
(15)
12
!
2`
E[ m2` ]
.
2` + 1
(16)
We emphasize that each of these equations for E[m ] is only valid when Theorem 2 is
satisfied for that particular value of m, and that the validity of one of these equations
does not imply the validity of any others corresponding to different m values.
Eq. (15) merits special comment. It indicates that if the total error variance in an
NSD quantizing system is input independent, then it always exceeds that of an SD
18
system (or a system described by the CMQ) by an amount equal to the variance of
the dither. This characteristic increase in the error power is not problematic in most
multi-bit applications, and the benefits of dithering typically far outweigh the slight
noise penalty.
Two corollaries to Theorem 2 follow.
Corollary 1 In an NSD quantizing system, if the condition of Eq. (13) is satisfied
for any given m, then for any choice of n
E[m xn ] = E[m ]E[xn ];
i.e., m and xn are uncorrelated.
Proof : We observe that if px (x) = (x x0 ) then
p () =
=
p,x(, x)dx
p|x(, x)(x x0 )dx
= p|x(, x0 ).
(17)
By Theorem 2, E[m ] is independent of the choice of px , and in particular it is independent of the choice of x0 when px (x) = (x x0 ), as above. Thus
Z
E[ x ] =
=
=
m xn p,x(, x)ddx
Z Z
= E[ ]E[xn ].
19
2
In particular, if E[] is independent of the distribution of x, then and x are uncorrelated in the usual mathematical sense:
E[x] = E[]E[x].
The second corollary is somewhat better known than Theorem 2 itself, but demands
satisfaction of a stronger condition [10, 30].
Corollary 2 In an NSD quantizing system, E[` ] is independent of the distribution
of the system input, x, for ` = 1, 2, . . . , m if and only if
(i)
P
k Z0
and
=0
i = 0, 1, 2, . . . , m 1.
!
`
X
`
i=0
We see that the `-th and all lower derivatives of G will all go to zero at u = k/,
k Z0 if the first ` 1 derivatives of P do. The only if direction is easily proven
using induction, but this requires more space than is justified here. The interested
reader is referred to [25].
2
In most practical applications, we are interested in dither signals which satisfy the
conditions of Corollary 2, and it turns out that the conditions of this corollary will be
20
of interest when we examine the statistics of the quantizer output (Section 2.4) and
the special nature of digital dither signals (Section 4).
2.3
We now begin an investigation into the joint statistics of temporally separated total
error values, corresponding to input samples separated in time, in order to derive
conclusions about the spectral characteristics of the total error sequence.
Consider two total error values, 1 and 2 , which are separated in time by 6= 0.
(In the special case where = 0, the analysis reduces to that of Section 2.2.) The corresponding system input values will be denoted as x1 and x2 , respectively. Employing
a derivation analogous to that of the Section 2.1 we find that
p
(1 ,2 )(x1 ,x2 )
(1 , 2 , x1 , x2 )
(1 + x1 k1 )(2 + x2 k2 )
k1 = k2 =
+k1
2
+k2
2
+k1
+k2
2
2
= [2 ? p1 ,2 ](1 , 2 )W (1 + x1 , 2 + x2 ),
where the convolution is two-dimensional, involving both 1 and 2 , and where
4
(1 , 2 ) = (1 ) (2 )
and
4
W (1 , 2 ) = W (1 )W (2 ).
p1 ,2 represents the joint pdf of the dither values, 1 and 2 , associated with the
inputs x1 and x2 , respectively.
21
Hence
p1 ,2 (1 , 2 ) =
=
p
(1 , 2 , x1 , x2 )px1 ,x2 (x1 , x2 )dx1 dx2
(1 ,2 ) (x1 ,x2 )
[2 ? p1 ,2 ](1 , 2 )[W ? px1 ,x2 ](1 , 2 ).
(18)
k1 = k2
k1
k2
sinc u1
sinc u2
P1 ,2
k1
k2
k1 k2
u1 , u2
Px1 ,x2 ,
.
(19)
No choice of dither pdf will allow Eq. (19) to be expressed as a product of two characteristic functions, one involving u1 alone and the other u2 alone, for arbitrary choices
of Px1 ,x2 . Thus 1 and 2 cannot be rendered statistically independent for arbitrary
joint input distributions. Let us therefore proceed to investigate the joint moments of
1 and 2 in the hope that we can exercise some control over them by an appropriate
choice of the dither statistics.
The (m1 , m2 )-th joint moment of the two signals of interest is given by:
4
1 m2
E[m
1 2 ] =
2 m2
m
1 2 p1 ,2 (1 , 2 )d1 d2
m1 +m2
j
2
1 ,m2 )
P(m
(0, 0)
1 ,2
(20)
where
4
1 ,m2 )
P(m
(u1 , u2 ) =
1 ,2
(m1 +m2 ) P1 ,2
(u1 , u2 ).
m2
1
um
1 u2
j
2
m1 +m2
Px1 ,x2
k1 = k2 =
22
k1 k2
k1 k2
1 ,m2 )
,
G(m
,
1 ,2
(21)
where
4
k1 k2
,
(k1 , k2 ) Z20 .
=0
(22)
The proof is completely analogous to that of Theorem 2. When Eq. (22) is satisfied,
we have
1 m2
E[m
1 2 ] =
j
2
m1 +m2
1 ,m2 )
G(m
(0, 0),
1 ,2
(23)
so that by explicitly performing the differentiation we can write an expression, analogous to Eq. (16), relating the joint moments of the total error to those of the dither:
1 m2
E[m
1 2 ]
m1
m2
bX
2 c bX
2 c
`1 =0 `2 =0
m1
2`1
m2
2`2
!
2(`1 +`2 )
We attach the caveat that satisfaction of Eq. (24) for some particular m1 and m2 does
not imply its satisfaction for any other values thereof.
If the dither process is iid so that 1 and 2 are statistically independent then
P1 ,2 (u1 , u2 ) = P (u1 )P (u2 ).
23
j m1 +m2 (m1 )
2)
=
G (0)G(m
(0)
2
m2
1
= E[m
1 ]E[2 ]
(25)
m2
m
m
m
1
so that m
1 and 2 are uncorrelated. In this case, of course, E[1 ] = E[2 ] = E[ ].
Hence:
Corollary 3 Any iid non-subtractive dither signal which satisfies the conditions of
Corollary 2 for m = max(m1 , m2 ) will ensure that, for two error values, 1 and 2 ,
separated in time by 6= 0,
m1
1 m2
]E[m2 ].
E[m
1 2 ] = E[
(26)
In this case E[m2 ] and E[m2 ] will be given by Eq. (16). In particular, for an iid
dither with zero mean we note that E[1 2 ] = 0.
In a digital system, the total error is a discrete-time signal, so that = kT where
T represents the sampling period and k Z. The autocorrelation function of such a
signal is defined to be E[1 2 ](k). The power spectral density (PSD) of a discrete-time
random process is equal by definition to the discrete-time Fourier transform (DTFT)
of its autocorrelation function, where we define the DTFT as
4
FDT [h](f ) = 2T
h(k)ej2f kT ,
(27)
k=
24
Using Eqs. (15) and (26) we find that, for an NSD quantizing system using iid
dither satisfying the conditions of Corollary 2 for m = 2, the autocorrelation function
of the error is
E[ 2 ] + , k = 0,
12
E[1 2 ](k) =
E [],
otherwise.
E[ 2 ], k = 0,
E 2 [], otherwise,
PSD (f ) = PSD (f ) +
2 T
6
so that the total error signal must be spectrally white since the dither is spectrally
white (apart from a dc component if the dither is not zero mean).
It is possible to derive conditions which ensure the satisfaction of Eq. (22) for
the case where m1 = m2 = 1, but which do not require statistical independence of
distinct dither values [24, 25]. This will allow the use of certain dither signals which
are notspectrally white.
Theorem 4 In an NSD system where all dither values are statistically independent
of all system input values,
E[1 2 ] = E[1 2 ]
(28)
for arbitrary input distributions if and only if the following three conditions are satisfied:
k1 k2
,
= 0
k1
P(0,1)
,
0
= 0
1 ,2
k2
P(1,0)
0,
= 0
1 ,2
P1 ,2
25
(k1 , k2 ) Z20 ,
(29)
k1 Z0 ,
(30)
k2 Z0 .
(31)
E[ 2 ] + , k = 0,
12
E[1 2 ](k) =
E[1 2 ],
(32)
otherwise.
This indicates that the power spectrum of the error will be identical to the power
spectrum of the dither, apart from a contribution due to the k = 0 case, manifested
as an additive constant present at all frequencies (i.e., a white spectral component
introduced by the properly dithered quantization operation). Hence, as before,
PSD (f ) = PSD (f ) +
2 T
,
6
(33)
except that now the dither PSD is not necessarily white. This will be illustrated by
the discussion of high-pass dither in Section 3.5.
26
2.4
2k1
2
2k+1
2 .
<w <
over this range where, since x and are statistically independent, pw is given by [34]
pw (w) = [p ? px ](w).
Thus we have
py (y) =
(y k)
k=
+k
2
+k
2
[p ? px ](w)dw
= [ ? p ? px ](y)W (y).
Taking the Fourier transform of this expression yields
Py (u) = [G (u)Px (u)] ? W 1 (u)
k=
k
k
u
Px u
(34)
and so
E[y m ] =
=
j
2
m
Py(m) (0)
X
m
X
k= r=0
!
m
r
j
2
r
G(r)
27
j
2
mr
Px(mr)
(35)
Now, if the first m derivatives of G (u) are zero at all non-zero multiples of 1/, then
Eq. (35) reduces to
m
E[y ] =
m
X
m
E[r ]E[xmr ],
r
r=0
(36)
where the expectation values of the total error are given in terms of the expectation
values of the dither by Eq. (16). By direct differentiation of G (u), the above condition
is easily shown to be equivalent to the condition of Corollary 2. For the special cases
m = 1 and m = 2 we note that
E[y] = E[x] + E[] = E[x] + E[]
E[y 2 ] = E[x2 ] + 2E[x]E[] + E[ 2 ] +
2
12
where Eqs. (14) and (15) have been substituted for the error moments.2
Proceeding similarly for the joint moments of output values y1 and y2 , separated
in time by 6= 0, we find that
E[y1m1 y2m2 ]
m1 X
m2
X
k1 = k2 = r1 =0 r2 =0
"
j
2
m1
r1
(m1 r1 )+(m2 r2 )
m2
r2
! "
j
2
r1 +r2
1 r1 ,m2 r2 )
Px(m
1 ,x2
1 ,r2 )
G(r
1 ,2
k1 k2
,
#
k1 k2
,
#
(37)
If the indicated partial derivatives of G1 ,2 are zero for all (k1 , k2 ) Z20 , ri =
1, 2, . . . , mi , i {1, 2}, then Eq. (37) reduces to
E[y1m1 y2m2 ]
m1 X
m2
X
r1 =0 r2 =0
2
m1
r1
m2
1 r1 m2 r2
E[r11 r22 ]E[xm
x2
],
1
r2
(38)
Note that the so-called dither averaged transfer characteristic, E[y|x], is given by
E[y|x] = E[Q(x + )|x] =
which is the convolution of the quantizer staircase with the dither pdf. For the m = 1 case this defines
the line y = x (see illustrations in [16, 17]).
28
where the joint moments of the total error are given in terms of those of the dither by
Eq. (24).
Beginning from Eq. (37) with m1 = m2 = 1 it is straightforward to show that if
the conditions of Theorem 4 are satisfied (i.e., Eqs. (29), (30) and (31)) then
E[y1 y2 ] = E[x1 x2 ] + E[1 2 ],
so that, with the aid of Eqs. (36) and (16), we find that the output has an autocorrelation function
E[x2 ] + 2E[x]E[] + E[ 2 ] + ,
12
E[y1 y2 ](k) =
E[x1 x2 ] + E[1 2 ],
k = 0,
(39)
otherwise.
Then the spectrum of the output is the sum of the input and dither spectra apart
from a white noise component, which is contributed by the k = 0 case of Eq. (39).
The latter component is comparable to the white quantization noise posited in the
CMQ. In particular, for a system using a zero-mean dither,
PSDy (f ) = PSDx (f ) + PSD (f ) +
2 T
.
6
We proceed to apply the above results to realizable quantizing systems using a variety
of dither signals.
3.1
Null Dither
29
the Fourier transform of which is equal to unity everywhere. Hence, by Eq. (12),
G (u) = sinc (u).
No derivatives of this function vanish at non-zero multiples of 1/, so no moments of
the total error (excepting the zeroth) will be independent of the input distribution. Of
course, it is not expected that they would be. We know that in the absence of dither,
the error is a deterministic function of the input. Indeed, the mean value of the error
as a function of the input is identical to the error function, q(x), shown in Fig. 2.
3.2
Rectangular-PDF Dither
Now consider a system using dither with a simple rectangular (i.e., uniform) pdf of
1 LSB peak-to-peak amplitude:
p () = (),
with a corresponding cf
P (u) = sinc (u).
Hence, from Eq. (12):
G (u) = sinc2 (u).
The first two derivatives of this function are plotted in Fig. 5. The first derivative
clearly satisfies the condition of going to zero at the regularly spaced points stipulated
by Eq. (13), while the second derivative does not (nor do higher derivatives). This
indicates that the first moment of the error signal is independent of the input, but that
its variance remains dependent. These conclusions are borne out by the accompanying
plots in Fig. 5 of the conditional moments
4
E[m |x] =
m p|x(, x)d,
30
1.0
1.5
0.8
First Moment [LSB]
2.0
1.0
0.5
0.0
-0.5
-1.0
-1.5
0.6
0.4
0.2
0.0
-0.2
-0.4
-0.6
-0.8
-2.0
-1.0
-4
-3
-2
-1
0
1
2
3
Frequency Variable, u [1/LSB]
-2
-1
0
1
System Input, x [LSB]
-2
-1
0
1
System Input, x [LSB]
0.40
0.35
Second Moment [LSB^2]
(a)
2
0
-2
-4
-6
0.30
0.25
0.20
0.15
0.10
0.05
-8
0.00
-4
-3
-2
-1
0
1
2
3
Frequency Variable, u [1/LSB]
(b)
Figure 5: Derivatives of G (u) (left) and conditional moments of the error (right) for a
(1)
quantizer using RPDF dither of 1 LSB peak-to-peak amplitude: (a) G (u) and E[|x]
(2)
(both in units of ), (b) G (u) and E[2 |x] (both in units of 2 ). The frequency
variable, u, is plotted in units of 1/ and the system input, x, in units of .
31
as computed using Eq. (7). The first moment, or mean error, is zero for all inputs,
indicating that the quantizer has been linearized by the use of this dither. The error
variance, on the other hand, is clearly signal dependent, so that the noise power in the
signal varies with the input. This is sometimes referred to as noise modulation and is
undesirable in audio or video signals.
If the dither is iid, then (as shown in Section 2.3), temporally separated error
values will be uncorrelated. Thus, short-time error spectra will appear flat but their
level will be input dependent.
3.3
Triangular-PDF Dither
The most straightforward means of generating dither signals with more complicated
pdfs is to simply sum two or more statistically independent RPDF random processes.
For instance, the sum of two such processes, 1 and 2 , each of 1 LSB peak-to-peak
amplitude, yields a dither with a triangular pdf (TPDF) of 2 LSB peak-to-peak amplitude, since the summation of statistically independent random processes convolves
their pdfs (see Fig. 6):
p () = [p1 ? p2 ]()
= [ ? ]().
(40)
32
p1 (x)
p2 (x)
1/
_ __
__
[p
1/
_ __
__
* p2 ](x)
1/
k
P(1)
= 0,
k Z0 ,
k Z.
A different proof of this result, using a direct method, was given in [16].
33
0.40
0.35
Second Moment [LSB^2]
2
0
-2
-4
-6
-8
0.30
0.25
0.20
0.15
0.10
0.05
-10
0.00
-4
-3
-2
-1
0
1
2
3
Frequency Variable, u [1/LSB]
-2
-1
0
1
System Input, x [LSB]
(2)
Figure 7: G (u) (left) and E[2 |x] (right) (both in units of 2 ) for a quantizer using
triangular-pdf dither of 2 LSB peak-to-peak amplitude. The frequency variable, u, is
plotted in units of 1/ and the system input, x, in units of .
since
P (0) =
ej2(0) p ()d = 1.
We conclude that the dither cf and its first derivative are completely specified at all
integer multiples of 1/. According to the Generalized Sampling Theorem [34], this is
sufficient to uniquely specify P (u) for all u if p () is -bandlimited (i.e., if p () = 0
for || ). Since the pdf of Eq. (40) is -bandlimited, and its corresponding cf
satisfies all the given conditions, it must be the unique pdf in question.
It remains to be shown that any dither pdf which is not thus bandlimited will
produce a greater error variance. Since this variance is assumed to be constant with
respect to the input, it is sufficient to show that this holds for a single input value.
We will do so for x = /2.
p|x (, x) for x = /2 is obtained from Eq. (8) using px (x) = (x
2)
(see
Eq. (17)). As is shown in Fig. 8(a) it consists of two equally weighted delta functions
at = /2 when triangular-pdf dither of 2 LSB peak-to-peak amplitude is employed.
34
p ( , x)
|x
1
2
p ( , x)
|x
x = /2
1
2
e-2
e-3
-
2
- 3
2
- 5
2
e-1
e1
-
2
x = /2
e2
3
2
e3
5
2
(b)
(a)
p|x ,
X
ei (2i 1)
i=1
+ ei + (2i 1)
(41)
We proceed by expressing the fundamental condition that the integral of this pdf must
equal unity:
(e1 + e1 ) +
(ei + ei ) = 1.
(42)
i=2
E[ |x = /2] =
=
X
i=1
2
(2i 1)
2
"
2
(ei + ei )
(e1 + e1 ) +
i=2
(43)
X
2
E[2 |x = /2] =
1+4
i(i 1)(ei + ei ) ,
4
i=2
"
which is always greater than 2 /4 since the ei s must be positive. We have thus shown
the following:
35
Theorem 5 The choice of zero-mean dither pdf which renders the first and second
moments of the total error independent of the input, such that the first moment is zero
and the second is minimized, is unique and is a triangular pdf of 2 LSB peak-to-peak
amplitude.
3.4
Theorem 6 A non-subtractive dither signal generated by the summation of n statistically independent RPDF random processes, renders E[` ] independent of the system
input distribution for ` = 0, 1, . . . , n, and results in a total error variance, for n 2,
of (n + 1)2 /12.
This must be the case since the use of n such dithers gives
G (u) = sincn+1 (u),
the first n derivatives of which will consist entirely of terms containing non-zero powers
of sinc (u). Since this function goes to zero at the required places, the first n moments
of the error will always be independent of the input. Higher derivatives will not share
this property [25]. Dithers of this form are sometimes referred to as nRPDF so that,
for instance, TPDF dither may also be referred to as 2RPDF.
It is important to note that using uniformly distributed processes of peak-to-peak
amplitude not equal to one LSB (or, rather, not equal to an integral number of LSBs)
will not render error moments independent of the input since the zeros of the associated
sinc functions will not fall at integral multiples of 1/ (see illustrations in [16]).
Finally, it is easily shown from the Generalized Sampling Theorem that the (n/2)bandlimited dither pdf which renders the first n moments of the total error independent
of the input is unique, and must therefore be the pdf of Theorem 6.
36
p (x)
p (x)
x
PRN
White
Rectangular-pdf
Input
z -1
High-Pass
Triangular-pdf
Output
3.5
High-Pass Dither
A very simple discrete-time noise generator capable of producing dither with a highpass spectrum is shown in Fig. 9. The system contains a pseudo-random number
generator, marked PRN, producing iid, uniformly distributed random numbers, and
a one-sample delay element marked z 1 . The output is the difference between the
pseudo-random number most recently generated by the PRN, n , and the previous
one, n1 ; that is, n = n n1 . The (first-order) pdf of the resulting dither sequence
is triangular (TPDF), since it results from the summation of two statistically independent RPDF sequences, albeit one of these is simply a delayed version of the other.
This means that all the beneficial effects of the TPDF dither discussed in Section 3.3
will also be associated with dither thus generated. Such high-pass TPDF dither may
be preferable in some audio applications since it is less audible than spectrally white
TPDF dither due to the ears reduced sensitivity at high frequencies (although these
dithers have equal variances of 2 /6). Similar comments apply regarding reduced error visibility in imaging applications. Furthermore, the use of high-pass TPDF dither
is more computationally efficient since it requires the calculation of only one new
37
RPDF random number per sample as compared to two when iid TPDF dither is used.
In order to investigate the spectral characteristics of the total error associated
with this sort of dither, we must derive an expression for p1 ,2 (1 , 2 ) as defined in
Section 2.3. Suppose that the sampling period of the system is T . For time lags
| | > T , the dither values are statistically independent so that
p1 ,2 (1 , 2 ) = p1 (1 )p2 (2 )
= [ ? ](1 )[ ? ](2 )
and
P1 ,2 (u1 , u2 ) = sinc2 (u1 )sinc2 (u2 ).
The non-trivial cases are those for = T . Consider two successive dither values
(i.e., = T ):
1 = 1 0
2 = 2 1 .
Then
p1 ,2 ,0 ,1 ,2 (1 , 2 , 0 , 1 , 2 )
= p1 |(2 ,0 ,1 ,2 ) (1 , 2 , 0 , 1 , 2 )p2 |0 ,1 ,2 (2 , 0 , 1 , 2 )p0 ,1 ,2 (0 , 1 , 2 )
= (1 1 + 0 )(2 2 + 1 )p0 (0 )p1 (1 )p2 (2 ).
Taking the Fourier transform of this expression with respect to all variables present
yields
P1 ,2 ,0 ,1 ,2 (u1 , u2 , w0 , w1 , w2 )
=
Desired marginal cfs can be obtained from a given joint cf by simply setting the
unwanted variables to zero, since
Px,y (u, 0) =
Thus we have
P1 ,2 (u1 , u2 ) = P0 (u1 )P1 (u1 u2 )P2 (u2 ).
Proceeding similarly for the case of = T , we find that
P1 ,2 (u1 , u2 ) = P0 (u2 )P1 (u2 u1 )P2 (u1 ).
For our purposes,
P0 (u) = P1 (u) = P2 (u) = sinc (u)
so that for both cases ( = T ):
P1 ,2 (u1 , u2 ) = sinc (u2 u1 )sinc (u1 )sinc (u2 ).
Finally, using = kT , k Z, we can write that
P1 ,2 (u1 , u2 ; k) =
k = 1,
(44)
|k| > 1.
It is straightforward to check that this joint cf satisfies all three conditions of Theorem 4.
Using Eq. (44) and the knowledge that TPDF dither has a variance of 2 /6, we
find using Eq. (20) that the autocorrelation function of the dither under consideration
is
1,
k = 0,
2
E[1 2 ](k) =
12 , k = 1,
0, otherwise.
39
2 T
[1 cos(2f T )],
3
1,
k = 0,
2
E[1 2 ](k) =
13 , k = 1,
0, otherwise,
2 T
[3 2 cos(2f T )].
6
This is simply the high-pass spectrum of the dither, plus a white quantization noise
component of 2 T /6 (which has a total power of 2 /12 up to the Nyquist frequency,
1
2T )
ci ni
i=
where is an iid nRPDF random process, the total error will be wide-sense stationary
and independent of the system input with a PSD given by
PSD (f ) = PSD (f ) +
under the following conditions:
40
2 T
6
1.
for each ` Z0 there exists an i such that of ci and ci+` one is zero and the
other is a non-zero integer,
and
2.
either is nRPDF with n 1 and there exist at least two distinct values of i
such that ci is a non-zero integer, or is nRPDF with n 2 and there exists at
least one value of i such that ci is a non-zero integer.
Digital Dither
(45)
p ()d =
p (`) = 1.
`=
For instance, p might be the pdf of a dither of order n, such as an nRPDF dither, in
which case it is straightforward to show using Poissons summation formula [35] that
p has the above normalization. In general, however, p need not correspond to a pdf
since it need not subtend unit area.
Taking the Fourier transform of Eq. (45) we find that
P (u) =
P ? W 1 (u)
`
=
P u
`=
(46)
where P (u) is the Fourier transform of p (). Note that even if P satisfies the
conditions of Corollary 2 (for some m), P will not, due to the modulation of P (u)
by the impulse train W 1 (u). Fortunately, we do not require that these conditions be
satisfied in a digital system, since the requirement that E[m |x] be constant for all
values of the system input is not of interest. Instead, we require only that the moments
be constant for a subset of all conceivable x values, namely {x|x = n, n Z}, which
42
includes all values that are representable in the digital system. Thus we assume that
the pdf of the system input can be expressed in the form
px (x) = px (x)W (x)
(47)
where px is a continuous function normalized such that the integral of Eq. (47) is unity.
Then
Px (u) = [Px ? W 1 ](u)
`
=
Px u
.
`=
(48)
G (u) =
=
sin(u)
P (u)
u
sin(u) X
k
P u
.
u k=
(49)
k=`=
k
k + 2L `
u
Px
so that
j m (m)
P (0)
2
!
m X
X
j
k
k + 2L `
(m)
=
Px
G
.
2
k=`=
E[m ] =
(50)
The only way that this quantity can be independent of Px is if we require that
(m)
G
=0
k
Z.
2L
(51)
That is, the indicated derivative must vanish for all integral values of k except those
which are integral multiples of 2L , the value of this derivative being immaterial in the
latter cases. In order to see that this is so, note that if a dither is chosen such that
Eq. (51) holds then many terms vanish from Eq. (50), leaving
m
E[ ] =
j
2
m X
G(m)
k=
X
Px
`=
`
.
Px
`=
= 1.
This leaves
E[m ] =
j
2
m X
G(m)
k=
k
,
(52)
which does not depend on the input distribution. The necessity of Eq. (51) follows
from the arbitrariness of Px (apart from its normalization). Furthermore, by inspection
Eq. (52) is precisely the m-th moment of a notional random variable with pdf
? p ()W (),
2L
although this is not, of course, the pdf of . Some algebraic manipulation of this
expression, exploiting the discrete-valued character of , reveals that it is equivalent
to the following:
W () ? p ().
2L
(53)
This may be regarded as the pdf of a notional random variable which is the sum of
the dither and an independent discrete-valued quantization noise.
Note also that in the limit as 0 (i.e., as L ) Eq. (51) becomes Eq. (13),
the condition of Theorem 2 for analog systems.
44
k= r=0
(54)
If P meets the conditions of Corollary 2, then all terms in Eq. (54) involving the
derivatives of P go to zero at the places required by Eq. (51) except for the single
(r = 0) term involving the m-th derivative. Fortunately, this term involves the zeroth
derivative of the leading sinc function, which goes to zero at all the required places.
This yields the following theorem:
Theorem 8 For a digital NSD system in which requantization is used to remove the
L least significant bits of binary data, E[` ] is independent of the input distribution
for ` = 1, 2, . . . , m, if a non-subtractive digital dither (with the same precision as the
input data) is applied for which
(i)
P
k Z0
and
=0
i = 0, 1, 2, . . . , m 1.
reflected in the appearance of a small non-zero mean error which, of course, will be
input independent if an appropriate dither pdf has been chosen.
To express the moments of the system output we impose the conditions of Theorem 8 upon Eq. (35), obtaining
m
E[y ] =
=
m
X
r=0
m
X
r=0
m
r
!
!
X
j r
k=
G(r)
"
j
2
mr
Px(mr)
#
m
E[r ]E[xmr ],
r
where we have observed from Eq. (48) that Px (u) is periodic with period 1/ so that
for any k Z
j
2
mr
Px(mr)
j
2
mr
sin(u)
P (u) ? W 1 (u).
u
46
This expression should be compared with Eq. (46). Note that if P satisfies the
conditions of the theorems, then so will the quantity
sin(u)
P (u).
u
Thus far, the behavior of the quantizer at step edges (i.e., when w =
2k1
2 ,
k Z)
has not been explicitly considered. This is not a problem if the signals in question are
continuous-valued. In this case the addition of dither will ensure that the quantizer
input resides at a quantizer-step edge with zero probability. On the other hand, if
digital signals are in use, the probability that the quantizer input resides at a step
edge is always greater than zero. In this instance it makes a considerable difference
to the quantizer output (and total error) whether the quantizer rounds up, down,
or stochastically (up or down with equal probability) at these edges. Technically, it
can be shown [25] that the above formalism yields correct predictions if a stochastic
quantizer is used.
The extension of the results to deterministic (i.e., non-stochastic) quantizers employs a simple trick. Consider, for instance, the consequences of choosing a quantizer
which always rounds up at step edges (a similar argument applies to quantizers which
round down). We note that if a (dc) virtual offset such that 0 < < is introduced
into the dither signal, the quantizer output is unaffected except that quantizer inputs residing at step edges are consistently rounded up. We can thus analyze digitally
dithered systems with deterministic requantizers using such a notional dc offset, which
is a purely mathematical device without physical counterpart. It can be shown [25]
that Theorem 8 holds precisely as before. Eq. (52) holds if the virtually offset dither
pdf
p () = p ( )W ( )
47
( )W () ? p ( + ).
2L
Conclusions
It is also worth noting that, since the dither is simply an additive signal which is independent of the system input, we are free to add it at any time prior to (re)quantization.
In particular, once a signal is properly dithered, other signals (which are statistically
independent of the dither) may be added to it and the resulting total signal will still
be properly dithered for (re)quantization purposes.
For audio signal processing purposes, there seems to be little point in rendering any
error moments other than the first and second independent of the input. Variations in
higher moments are believed to be inaudible and this has been corroborated by a large
number of psycho-acoustic tests conducted by the authors and others [13, 21]. These
tests involved listening to a large variety of signals (sinusoids, sinusoidal chirps, slow
ramps, various periodically switched inputs, piano and orchestral music, etc.) which
had been very coarsely requantized (from 16 bits to 8 bits) in order to render the
requantization error essentially independent of low-level non-linearities in the digitalto-analog conversion system used for listening purposes. In addition, the corresponding
error signals (output minus input) were used in listening tests in order to check for
any vestiges of audible dependence on the input. Using undithered quantizers resulted
in clearly audible distortion and noise modulation in the output and error signals.
Rectangular-pdf dither of 1 LSB peak-to-peak amplitude eliminated all distortion, but
the residual noise level was found to vary audibly in an input-dependent fashion. When
triangular-pdf dither of 2 LSB peak-to-peak amplitude (either white or high-pass)
was employed, no instance was found in which the error was audibly distinguishable
from a steady random noise entirely unrelated to the input. Admittedly, these tests
were informal, and there remains a need for formal psycho-acoustic tests of this sort
involving many participants under carefully controlled conditions.
We recommend the use of spectrally-white triangular-pdf (TPDF) dither of 2 LSB
peak-to-peak amplitude for most audio applications requiring non-subtractively dithered
49
51
Acknowledgments
Stanley P. Lipshitz and John Vanderkooy have been supported by operating grants
from the Natural Sciences and Engineering Research Council of Canada.
References
[1] Roberts, L.G., Picture Coding Using Pseudo-Random Noise, IRE Trans. Inform. Theory, vol. IT-8, pp. 145154, (1962 Feb.).
[2] Jayant, N.S. and L.R. Rabiner, The Application of Dither to the Quantization
of Speech Signals, Bell Syst. Tech. J., vol. 51, pp. 12931304 (1972 JulyAug.).
[3] Schuchman, L., Dither Signals and Their Effect on Quantization Noise, IEEE
Trans. Commun. Technol., vol. COM-12, pp. 162165 (1964 Dec.).
[4] Sherwood, D.T., Some Theorems on Quantization and An Example Using
Dither, Conference Record, 19th Asilomar Conference on Circuits, Systems, and
Computers, Pacific Grove, CA (1985 Nov).
[5] Widrow, B., A Study of Rough Amplitude Quantization by Means of Nyquist
Sampling Theory, Ph.D Thesis, Department of Electrical Engineering, Massachusetts Institute of Technology, Cambridge, MA, (1956 Jun.).
[6] Widrow, B., A Study of Rough Amplitude Quantization by Means of Nyquist
Sampling Theory, IRE Trans. Circuit Theory, vol. PGCT-3, no. 4, pp. 266276,
(1956 Dec.).
[7] Widrow, B., Statistical Analysis of Amplitude-Quantized Sampled-Data Systems, Trans. Amer. Inst. Elec. Eng., Pt. II, Applications and Industry, vol. 79,
pp.555568 (1961 Jan.).
52
[8] Sripad, A.B., and D.L. Snyder, A Necessary and Sufficient Condition for Quantization Errors to Be Uniform and White, IEEE Trans. Acoust., Speech, Signal
Process., vol. ASSP-25, pp. 442448 (1977 Oct.).
[9] Gray, R.N., Quantization Noise Spectra, IEEE Trans. Info. Theory., pp. 1220
1244 (1990 Nov.).
[10] Wright, J.N., unpublished manuscripts (1979 Jun.Aug.).
[11] Wright, J.N, private communication, (1991 Apr.).
[12] Stockham, T.G., private communication (1988).
[13] Brinton, L.K., Nonsubtractive Dither, M.Sc. Thesis, Dept. of Elec. Eng., Univ.
of Utah, Salt Lake City, UT (1984 Aug.).
[14] Gray, R.M., and T.G. Stockham, Dithered Quantizers, IEEE Trans. Inform.
Theory, vol. 39, pp. 805811 (1993 May).
[15] Vanderkooy, J., and S.P. Lipshitz, Resolution Below the Least Significant Bit
in Digital Systems with Dither, J. Audio Eng. Soc., vol. 32, pp. 106113
(1984 Mar.); correction ibid., p. 889 (1984 Nov.).
[16] Lipshitz, S.P., and J. Vanderkooy, Digital Dither, presented at the 81st Convention of the Audio Engineering Society, J. Audio Eng. Soc. (Abstracts), vol. 34,
p. 1030 (1986 Dec.), preprint 2412.
[17] Vanderkooy, J., and S.P. Lipshitz, Dither in Digital Audio, J. Audio Eng. Soc.,
vol. 35, pp. 966975 (1987 Dec.).
53
[18] Vanderkooy, J., and S.P. Lipshitz, Digital Dither: Signal Processing with Resolution Far Below the Least Significant Bit, Proc. of the AES 7th International
Conference: Audio in Digital Times, Toronto, Canada, pp. 8796 (1989 May).
[19] Lipshitz, S.P., and J. Vanderkooy, High-Pass Dither, presented at the 4th Regional Convention of the Audio Engineering Society, Tokyo (1989 Jun.); in Collected Preprints (AES Japan Section, Tokyo, 1989), pp. 7275.
[20] Wannamaker, R.A., S.P. Lipshitz and J. Vanderkooy, Dithering to Eliminate
Quantization Distortion, Proc. Annual Meeting Can. Acoustical Assoc., Halifax,
NS, Canada, pp. 7886 (1989 Oct.).
[21] Wannamaker, R.A., Dither and Noise Shaping in Audio Applications, M.Sc.
Thesis, Dept. of Physics, Univ. of Waterloo, Waterloo, ON, Canada, (1990 Dec.).
[22] Lipshitz, S.P., R.A. Wannamaker, J. Vanderkooy, and J.N. Wright, NonSubtractive Dither, Proc. of the 1991 IEEE Workshop on Applications of Signal
Processing to Audio and Acoustics, New Paltz, NY (1991 Oct.), Paper No. 6.2.
[23] Wannamaker, R.A., and S.P. Lipshitz, Time Domain Behavior of Dithered Quantizers, presented at the 93rd Convention of the Audio Engineering Society, San
Francisco, CA (1992 Oct.), preprint 3418.
[24] Wannamaker, R.A., Subtractive and Non-Subtractive Dithering: A Comparative
Analysis, presented at the 97th Convention of the Audio Engineering Society,
San Francisco, CA (1994 Nov.), preprint 3920.
[25] Wannamaker, R.A., The Mathematical Theory of Dithered Quantization,
Ph.D. Thesis, Dept. of Applied Mathematics, Univ. of Waterloo, Waterloo, ON,
Canada, (1997 Jun.).
54
[26] Lipshitz, S.P., R.A. Wannamaker and J. Vanderkooy, Quantization and Dither:
A Theoretical Survey, J. Audio Eng. Soc., vol. 40, pp. 355375 (1992 May).
[27] Lipshitz, S.P., R.A. Wannamaker, and J. Vanderkooy, Dithered Noise Shapers
and Recursive Digital Filters, presented at the 94th Convention of the Audio
Engineering Society, Berlin, Germany (1993 Mar.), preprint 3515.
[28] Wannamaker, R.A., and S.P. Lipshitz, Dithered Quantizers With and Without
Feedback, Proc. of the 1993 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics, New Paltz, NY (1993 Oct.).
[29] Wannamaker, R.A., Efficient Generation of Multichannel Dither Signals, presented at the 103rd Convention of the Audio Engineering Society, New York, NY
(1997 Sept.), preprint 4533.
[30] Gray, R.M., private communication, (1991 Apr.).
[31] Jayant, N.S., and P. Noll, Digital Coding of Waveforms, Prentice Hall, Englewood
Cliffs, NJ, (1984).
[32] Lukacs, E., Characteristic Functions, Charles Griffin & Co., London, UK (1960).
[33] Kawata, T., Fourier Analysis in Probability Theory, Academic Press, NY, NY
(1972).
[34] Papoulis, A., Probability, Random Variables, and Stochastic Processes, 2nd ed.,
McGraw-Hill, New York, NY, (1984).
[35] Papoulis, A., The Fourier Integral and Its Applications, McGraw-Hill, NY, (1962).
[36] Knuth, D., The Art of Computer Programming, vol. 2, 2nd ed. Addison-Wesley,
Reading, MA (1981).
55
[37] Lipshitz, S.P., J. Vanderkooy and R.A. Wannamaker, Minimally Audible Noise
Shaping, J. Audio Eng. Soc., vol. 39, pp. 836852 (1991 Nov.)
[38] Wannamaker, R.A., Psychoacoustically Optimal Noise Shaping, J. Audio Eng.
Soc., vol. 40, pp. 611620 (1992 July/Aug.).
[39] Norsworthy, S.R., R. Schreier and G.C. Temes, Delta-Sigma Data Converters:
Theory, Design and Simulation, IEEE Press, Piscataway, NJ (1997).
56